术中脑肿瘤分类:激光诱导荧光光谱与机器学习
本研究在手术室使用无损激光内源性荧光光谱设备TumorID扫描46例患者的离体标本,包括胶质瘤8例、脑膜瘤10例、垂体腺瘤23例及癫痫手术切除的非肿瘤组织5例,采用支持向量机算法进行分类。模型多分类受试者工作特征曲线下面积为0.809±0.002,中性卟啉的曲线下面积具有统计学显著性(p<0.001)且对模型输出影响最大。
要点
- TumorID仅需0.5秒采集数据,无需染料或组织操作,使用100毫瓦405纳米激光且不损伤组织。
- 在46例离体标本中,机器学习模型多分类AUC为0.809±0.002。
- 中性卟啉的AUC具有统计学显著性(p<0.001),对模型输出影响最大。
核心结论
- 这项初步离体临床研究显示TumorID能快速区分和分类不同病理及周围脑组织,其配置可易于转化为在体扫描。
- 作者认为该分类能力可使TumorID通过快速术中组织诊断和边界描绘来辅助手术决策,可能通过更知情和完整的切除改善患者结局。
研究设计与样本
- 研究使用无损激光内源性荧光光谱设备TumorID,在手术室扫描46例患者的离体标本,患者平均年龄52岁。
- 标本包括胶质瘤8例、脑膜瘤10例、垂体腺瘤23例及癫痫手术切除的非肿瘤组织5例。
- 训练支持向量机算法区分这些病变并近实时分类,统计显著性通过广义估计方程基于已知荧光团发射区域面积确定。
主要发现
- 机器学习模型显示高度分类能力,多分类受试者工作特征曲线下面积为0.809±0.002。
- 中性卟啉的曲线下面积具有统计学显著性(p<0.001),且对模型输出影响最大。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: To optimize neurosurgical tumor resection, tissue types and borders must be appropriately identified. Authors of this study established the use of a nondestructive laser-based endogenous fluorescence spectroscopy device, "TumorID," to almost immediately classify a specimen as glioma, meningioma, pituitary adenoma, or nonneoplastic tissue in the operating room, utilizing a machine learning algorithm. METHODS: TumorID requires only 0.5 seconds to collect data, without the need for any dyes or tissue manipulation, and utilizes a 100-mW, 405-nm laser that does not damage the tissue. The system was used in the operating room to scan ex vivo specimens from 46 patients (mean age 52 years) with glioma (8 patients), meningioma (10 patients), pituitary adenoma (23 patients), and nonneoplastic tissue resected during an epilepsy operation (5 patients). A support vector machine algorithm was trained to distinguish between these lesions and classify them in near real time. Statistical significance was determined through a generalized estimating equation on the area under the known fluorophore emission regions for free reduced nicotinamide adenine dinucleotide (NADH), bound NADH, flavin adenine dinucleotide, and neutral porphyrins. RESULTS: Ultimately, the machine learning model showed a high degree of classification power with a multiclass area under the receiver operating characteristic curve of 0.809 ± 0.002. The areas under the curve for neutral porphyrins were found to be statistically significant (p < 0.001) and to have the largest impact on model output. CONCLUSIONS: This initial ex vivo clinical study demonstrated the ability of TumorID to rapidly differentiate and classify various pathologies and surrounding brain in a configuration that can be easily translated to scan in vivo. This classification power could allow TumorID to augment surgical decision-making by enabling rapid intraoperative tissue diagnostics and border delineation, potentially improving patient outcomes by allowing for a more informed and complete resection.
引用
- Zachem TJ, Sperber JE, Chen SF, et al. Intraoperative brain tumor classification via laser-induced fluorescence spectroscopy and machine learning. J Neurosurg. 2025. PMID: 40184679. doi:10.3171/2024.12.JNS242041
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。